Neural Object Relational Models (Norm)¶
When human defines a concept, we compose it with other concepts. When the concept is not accurate, human critic and modify the logic of the composition. However, that is not what the state of the arts AI technology, deep learning practices. Deep neural networks considers the concept as a set of numerical parameters to be optimized with respect to a set of data and an objective function. This black-box approach is difficult for regular human experts to interpret and modify. It requires an experienced neural network architect to fine-tune the parameters constantly.
For example, JayWalk is a new concept that we need to detect for Autonomous Driving Vehicles. The standard procedure is to collect a set of positive and negative images, then train a model to classify this concept. The challenge is that it is difficult to collect positive examples for a complicated concept due to the long tail distribution and the trained model will be less accurate if the data is not enough. However, if we compose the concept based on other well-trained concepts, the chance to obtain a high quality model will be increased significantly.
JayWalk(p: Person, r: Road) = WalkAcross(p, r) & On(p, z) & Part(r, ?z) & !ZebraCross(z);
If some concepts contain errors that accumulate due to the complex compositions, Norm can alleviate this brittleness effectively by adapting the parameters over a small set of examples. The entire logic program is compiled to a neural network and the power of transfer learning is leveraged.
Representing the AI model in terms of logic forms facilitates the white-box machine learning approach. Domain experts can understand the logical explanation of the model and can argue with the model by looking into the counter-examples. Particularly, domain experts can append the differential logic to the program and test them out. These explanatory and explorative debugging capabilities turn the AI model development into an interactive and collaborative process that fits into most of research agenda in many fields.
In the end, logic program is the closest formal language to human language. Having a natural language parser will enable many more people to benefit the AI development who do not know how to write in a formal language. It will open a door to the Artificial General Intelligence that is capable of human-like reasoning.
- Logic Programming
- Neural Networks
- Human Interactions